1 // sgmm2bin/sgmm2-latgen-faster-parallel.cc
3 // Copyright 2009-2013 Saarland University; Microsoft Corporation;
4 // Johns Hopkins University (author: Daniel Povey)
5 // 2014 Guoguo Chen
7 // See ../../COPYING for clarification regarding multiple authors
8 //
9 // Licensed under the Apache License, Version 2.0 (the "License");
10 // you may not use this file except in compliance with the License.
11 // You may obtain a copy of the License at
12 //
13 // http://www.apache.org/licenses/LICENSE-2.0
14 //
15 // THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
16 // KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
17 // WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
18 // MERCHANTABLITY OR NON-INFRINGEMENT.
19 // See the Apache 2 License for the specific language governing permissions and
20 // limitations under the License.
22 #include <string>
23 using std::string;
25 #include "base/kaldi-common.h"
26 #include "util/common-utils.h"
27 #include "sgmm2/am-sgmm2.h"
28 #include "hmm/transition-model.h"
29 #include "fstext/fstext-lib.h"
30 #include "decoder/decoder-wrappers.h"
31 #include "sgmm2/decodable-am-sgmm2.h"
32 #include "util/kaldi-thread.h"
33 #include "base/timer.h"
35 namespace kaldi {
37 // the reference arguments at the beginning are not const as the style guide
38 // requires, but are best viewed as inputs.
39 void ProcessUtterance(const AmSgmm2 &am_sgmm,
40 const TransitionModel &trans_model,
41 double log_prune,
42 double acoustic_scale,
43 const Matrix<BaseFloat> &features,
44 RandomAccessInt32VectorVectorReader &gselect_reader,
45 RandomAccessBaseFloatVectorReaderMapped &spkvecs_reader,
46 const fst::SymbolTable *word_syms,
47 const std::string &utt,
48 bool determinize,
49 bool allow_partial,
50 Int32VectorWriter *alignments_writer,
51 Int32VectorWriter *words_writer,
52 CompactLatticeWriter *compact_lattice_writer,
53 LatticeWriter *lattice_writer,
54 LatticeFasterDecoder *decoder, // Takes ownership of this.
55 double *like_sum,
56 int64 *frame_sum,
57 int32 *num_done,
58 int32 *num_err,
59 TaskSequencer<DecodeUtteranceLatticeFasterClass> *sequencer) {
60 using fst::Fst;
61 using std::vector;
63 Sgmm2PerSpkDerivedVars *spk_vars = new Sgmm2PerSpkDerivedVars; // decodable
64 // will take ownership.
65 if (spkvecs_reader.IsOpen()) {
66 if (spkvecs_reader.HasKey(utt)) {
67 spk_vars->SetSpeakerVector(spkvecs_reader.Value(utt));
68 am_sgmm.ComputePerSpkDerivedVars(spk_vars);
69 } else {
70 KALDI_WARN << "Cannot find speaker vector for " << utt << ", not decoding this utterance";
71 delete spk_vars;
72 (*num_err)++;
73 return;
74 }
75 }
76 if (!gselect_reader.HasKey(utt) ||
77 gselect_reader.Value(utt).size() != features.NumRows()) {
78 KALDI_WARN << "No Gaussian-selection info available for utterance "
79 << utt << " (or wrong size)";
80 }
82 // decodable will take ownership.
83 vector<vector<int32> > *gselect = new std::vector<vector<int32> >(
84 gselect_reader.Value(utt));
86 Matrix<BaseFloat> *new_feats = new Matrix<BaseFloat>(features); // decodable
87 // will take ownership of this.
89 // This takes ownership of new_feats, gselect, and spk_vars
90 DecodableAmSgmm2Scaled *sgmm_decodable = new DecodableAmSgmm2Scaled(
91 am_sgmm, trans_model, new_feats, gselect,
92 spk_vars, log_prune, acoustic_scale);
94 // takes ownership of decoder and sgmm_decodable.
95 DecodeUtteranceLatticeFasterClass *task =
96 new DecodeUtteranceLatticeFasterClass(
97 decoder, sgmm_decodable, trans_model, word_syms, utt, acoustic_scale,
98 determinize, allow_partial, alignments_writer, words_writer,
99 compact_lattice_writer, lattice_writer, like_sum, frame_sum, num_done,
100 num_err, NULL);
102 sequencer->Run(task); // takes ownership.
103 }
105 } // end namespace kaldi
107 int main(int argc, char *argv[]) {
108 try {
109 using namespace kaldi;
110 typedef kaldi::int32 int32;
111 using fst::SymbolTable;
112 using fst::Fst;
113 using fst::VectorFst;
114 using fst::StdArc;
116 const char *usage =
117 "Decode features using SGMM-based model. This version accepts the --num-threads\n"
118 "option but otherwise behaves identically to sgmm2-latgen-faster\n"
119 "Usage: sgmm2-latgen-faster-parallel [options] <model-in> (<fst-in>|<fsts-rspecifier>) "
120 "<features-rspecifier> <lattices-wspecifier> [<words-wspecifier> [<alignments-wspecifier>] ]\n";
121 ParseOptions po(usage);
122 BaseFloat acoustic_scale = 0.1;
123 bool allow_partial = false;
124 BaseFloat log_prune = 5.0;
125 string word_syms_filename, gselect_rspecifier, spkvecs_rspecifier,
126 utt2spk_rspecifier;
128 LatticeFasterDecoderConfig decoder_opts;
129 TaskSequencerConfig sequencer_config; // has --num-threads option
130 decoder_opts.Register(&po);
131 sequencer_config.Register(&po);
133 po.Register("acoustic-scale", &acoustic_scale,
134 "Scaling factor for acoustic likelihoods");
135 po.Register("log-prune", &log_prune,
136 "Pruning beam used to reduce number of exp() evaluations.");
137 po.Register("word-symbol-table", &word_syms_filename,
138 "Symbol table for words [for debug output]");
139 po.Register("allow-partial", &allow_partial,
140 "Produce output even when final state was not reached");
141 po.Register("gselect", &gselect_rspecifier,
142 "rspecifier for precomputed per-frame Gaussian indices.");
143 po.Register("spk-vecs", &spkvecs_rspecifier,
144 "rspecifier for speaker vectors");
145 po.Register("utt2spk", &utt2spk_rspecifier,
146 "rspecifier for utterance to speaker map");
147 po.Read(argc, argv);
149 if (po.NumArgs() < 4 || po.NumArgs() > 6) {
150 po.PrintUsage();
151 exit(1);
152 }
154 if (gselect_rspecifier == "")
155 KALDI_ERR << "--gselect option is required.";
157 std::string model_in_filename = po.GetArg(1),
158 fst_in_str = po.GetArg(2),
159 feature_rspecifier = po.GetArg(3),
160 lattice_wspecifier = po.GetArg(4),
161 words_wspecifier = po.GetOptArg(5),
162 alignment_wspecifier = po.GetOptArg(6);
164 double tot_like = 0.0;
165 kaldi::int64 frame_count = 0;
166 int num_done = 0, num_err = 0;
167 Timer timer;
168 Fst<StdArc> *decode_fst = NULL;
169 fst::SymbolTable *word_syms = NULL;
171 TaskSequencer<DecodeUtteranceLatticeFasterClass> sequencer(
172 sequencer_config);
173 TransitionModel trans_model;
174 kaldi::AmSgmm2 am_sgmm;
175 {
176 bool binary;
177 Input ki(model_in_filename, &binary);
178 trans_model.Read(ki.Stream(), binary);
179 am_sgmm.Read(ki.Stream(), binary);
180 }
182 CompactLatticeWriter compact_lattice_writer;
183 LatticeWriter lattice_writer;
185 bool determinize = decoder_opts.determinize_lattice;
186 if (! (determinize ? compact_lattice_writer.Open(lattice_wspecifier)
187 : lattice_writer.Open(lattice_wspecifier)))
188 KALDI_ERR << "Could not open table for writing lattices: "
189 << lattice_wspecifier;
191 Int32VectorWriter words_writer(words_wspecifier);
192 Int32VectorWriter alignment_writer(alignment_wspecifier);
194 if (word_syms_filename != "")
195 if (!(word_syms = fst::SymbolTable::ReadText(word_syms_filename)))
196 KALDI_ERR << "Could not read symbol table from file "
197 << word_syms_filename;
199 RandomAccessInt32VectorVectorReader gselect_reader(gselect_rspecifier);
200 RandomAccessBaseFloatVectorReaderMapped spkvecs_reader(spkvecs_rspecifier,
201 utt2spk_rspecifier);
203 if (ClassifyRspecifier(fst_in_str, NULL, NULL) == kNoRspecifier) { // a single FST.
204 SequentialBaseFloatMatrixReader feature_reader(feature_rspecifier);
205 // It's important that we initialize decode_fst after feature_reader, as it
206 // can prevent crashes on systems installed without enough virtual memory.
207 // It has to do with what happens on UNIX systems if you call fork() on a
208 // large process: the page-table entries are duplicated, which requires a
209 // lot of virtual memory.
210 decode_fst = fst::ReadFstKaldiGeneric(fst_in_str);
211 timer.Reset(); // exclude graph loading time.
213 {
214 for (; !feature_reader.Done(); feature_reader.Next()) {
215 string utt = feature_reader.Key();
216 const Matrix<BaseFloat> &features(feature_reader.Value());
217 if (features.NumRows() == 0) {
218 KALDI_WARN << "Zero-length utterance: " << utt;
219 num_err++;
220 continue;
221 }
223 // ProcessUtterance will take ownership of this.
224 LatticeFasterDecoder *decoder = new LatticeFasterDecoder(
225 *decode_fst, decoder_opts);
227 ProcessUtterance(am_sgmm, trans_model, log_prune, acoustic_scale,
228 features, gselect_reader, spkvecs_reader, word_syms,
229 utt, determinize, allow_partial,
230 &alignment_writer, &words_writer, &compact_lattice_writer,
231 &lattice_writer, decoder, &tot_like, &frame_count,
232 &num_done, &num_err, &sequencer);
233 }
234 }
235 } else { // We have different FSTs for different utterances.
236 SequentialTableReader<fst::VectorFstHolder> fst_reader(fst_in_str);
237 RandomAccessBaseFloatMatrixReader feature_reader(feature_rspecifier);
238 for (; !fst_reader.Done(); fst_reader.Next()) {
239 std::string utt = fst_reader.Key();
240 if (!feature_reader.HasKey(utt)) {
241 KALDI_WARN << "Not decoding utterance " << utt
242 << " because no features available.";
243 num_err++;
244 continue;
245 }
246 const Matrix<BaseFloat> &features = feature_reader.Value(utt);
247 if (features.NumRows() == 0) {
248 KALDI_WARN << "Zero-length utterance: " << utt;
249 num_err++;
250 continue;
251 }
252 VectorFst<StdArc> *fst = fst_reader.Value().Copy(); // Note: this does
253 // a shallow copy because OpenFst is "smart" about these things and
254 // does reference counting. The constructor of LatticeFasterDecoder
255 // takes ownership of this FST (note: LatticeFasterDecoder has 2
256 // constructors, one of which takes ownership and one of which does not).
257 LatticeFasterDecoder *decoder = new LatticeFasterDecoder(decoder_opts,
258 fst);
260 // ProcessUtterance takes ownership of "decoder".
261 ProcessUtterance(am_sgmm, trans_model, log_prune, acoustic_scale,
262 features, gselect_reader, spkvecs_reader, word_syms,
263 utt, determinize, allow_partial,
264 &alignment_writer, &words_writer, &compact_lattice_writer,
265 &lattice_writer, decoder, &tot_like, &frame_count,
266 &num_done, &num_err, &sequencer);
267 }
268 }
269 sequencer.Wait(); // Wait till all tasks are done.
271 delete decode_fst;
272 delete word_syms;
274 double elapsed = timer.Elapsed();
275 KALDI_LOG << "Decoded with " << sequencer_config.num_threads << " threads.";
276 KALDI_LOG << "Time taken [excluding initialization] "<< elapsed
277 << "s: real-time factor per thread assuming 100 frames/sec is "
278 << (sequencer_config.num_threads * elapsed * 100.0 / frame_count);
279 KALDI_LOG << "Done " << num_done << " utterances, failed for "
280 << num_err;
281 KALDI_LOG << "Overall log-likelihood per frame = " << (tot_like/frame_count)
282 << " over " << frame_count << " frames.";
284 return (num_done != 0 ? 0 : 1);
285 } catch(const std::exception &e) {
286 std::cerr << e.what();
287 return -1;
288 }
289 }